diff --git a/python/llm/example/GPU/Pipeline-Parallel-Inference/README.md b/python/llm/example/GPU/Pipeline-Parallel-Inference/README.md
index 1d099da3..19d14cea 100644
--- a/python/llm/example/GPU/Pipeline-Parallel-Inference/README.md
+++ b/python/llm/example/GPU/Pipeline-Parallel-Inference/README.md
@@ -17,7 +17,17 @@ To run this example with IPEX-LLM on Intel GPUs, we have some recommended requir
- [baichuan-inc/Baichuan2-13B-Chat](./run_baichuan2_arc_2_card.sh)
- [microsoft/Phi-3-mini-4k-instruct](./run_phi3_arc_2_card.sh)
- [microsoft/Phi-3-medium-4k-instruct](./run_phi3_arc_2_card.sh)
-
+- [mistralai/Mistral-7B-v0.1](./run_mistral_arc_2_card.sh)
+- [mistralai/Mixtral-8x7B-Instruct-v0.1](./run_mistral_arc_2_card.sh)
+- [01-ai/Yi-6B-Chat](./run_yi_arc_2_card.sh)
+- [01-ai/Yi-34B-Chat](./run_yi_arc_2_card.sh)
+- [codellama/CodeLlama-7b-Instruct-hf](./run_codellama_arc_2_card.sh)
+- [codellama/CodeLlama-13b-Instruct-hf](./run_codellama_arc_2_card.sh)
+- [codellama/CodeLlama-34b-Instruct-hf](./run_codellama_arc_2_card.sh)
+- [upstage/SOLAR-10.7B-Instruct-v1.0](./run_solar_arc_2_card.sh)
+- [lmsys/vicuna-7b-v1.3](./run_vicuna_arc_2_card.sh)
+- [lmsys/vicuna-13b-v1.3](./run_vicuna_arc_2_card.sh)
+- [lmsys/vicuna-33b-v1.3](./run_vicuna_arc_2_card.sh)
## Example: Run pipeline parallel inference on multiple GPUs
@@ -95,7 +105,6 @@ bash run_chatglm_arc_2_card.sh
You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Baichuan2 to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
```bash
-pip install transformers==4.37.0
bash run_baichuan2_arc_2_card.sh
```
@@ -117,6 +126,83 @@ bash run_phi3_arc_2_card.sh
+
+
+
+ Show Mistral/Mixtral example
+
+#### Run Mistral-7B-v0.1 / Mixtral-8x7B-Instruct-v0.1 on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Mistral / Mixtral to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+pip install transformers==4.37.0
+bash run_mistral_arc_2_card.sh
+```
+
+
+
+
+
+
+ Show Yi example
+
+#### Run Yi-6B-Chat / Yi-34B-Chat on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Yi to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+bash run_yi_arc_2_card.sh
+```
+
+
+
+
+
+
+ Show Codellama example
+
+#### Run CodeLlama-7b-Instruct-hf / CodeLlama-13b-Instruct-hf / CodeLlama-34b-Instruct-hf on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Codellama to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+bash run_codellama_arc_2_card.sh
+```
+
+
+
+
+
+
+ Show Solar example
+
+#### Run SOLAR-10.7B-Instruct-v1.0 on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Solar to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+bash run_solar_arc_2_card.sh
+```
+
+
+
+
+
+
+ Show Vicuna example
+
+#### Run vicuna-7b-v1.3 / vicuna-13b-v1.3 / vicuna-33b-v1.3 on two Intel Arc A770
+
+You could specify `--repo-id-or-model-path` in the test script to be the huggingface repo id for Vicuna to be downloaded, or the path to the huggingface checkpoint folder. Besides, you could change `NUM_GPUS` to the number of GPUs you have on your machine.
+
+```bash
+bash run_vicuna_arc_2_card.sh
+```
+
+
+
+
### 3. Sample Output
#### [meta-llama/Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf)
```log
diff --git a/python/llm/example/GPU/Pipeline-Parallel-Inference/run_codellama_arc_2_card.sh b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_codellama_arc_2_card.sh
new file mode 100644
index 00000000..3de4c1ca
--- /dev/null
+++ b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_codellama_arc_2_card.sh
@@ -0,0 +1,41 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+ export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run CodeLlama-7b-Instruct-hf
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+ generate.py --repo-id-or-model-path 'codellama/CodeLlama-7b-Instruct-hf' --gpu-num $NUM_GPUS
+
+# To run CodeLlama-13b-Instruct-hf
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+# generate.py --repo-id-or-model-path 'codellama/CodeLlama-13b-Instruct-hf' --gpu-num $NUM_GPUS
+
+# To run CodeLlama-34b-Instruct-hf
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+# generate.py --repo-id-or-model-path 'codellama/CodeLlama-34b-Instruct-hf' --gpu-num $NUM_GPUS
diff --git a/python/llm/example/GPU/Pipeline-Parallel-Inference/run_mistral_arc_2_card.sh b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_mistral_arc_2_card.sh
new file mode 100644
index 00000000..0eb36a42
--- /dev/null
+++ b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_mistral_arc_2_card.sh
@@ -0,0 +1,37 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+ export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run Mistral-7B-v0.1
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+ generate.py --repo-id-or-model-path 'mistralai/Mistral-7B-v0.1' --gpu-num $NUM_GPUS
+
+# To run Mixtral-8x7B-Instruct-v0.1
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+# generate.py --repo-id-or-model-path 'mistralai/Mixtral-8x7B-Instruct-v0.1' --gpu-num $NUM_GPUS
diff --git a/python/llm/example/GPU/Pipeline-Parallel-Inference/run_solar_arc_2_card.sh b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_solar_arc_2_card.sh
new file mode 100644
index 00000000..76781745
--- /dev/null
+++ b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_solar_arc_2_card.sh
@@ -0,0 +1,33 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+ export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run SOLAR-10.7B-Instruct-v1.0
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+ generate.py --repo-id-or-model-path 'upstage/SOLAR-10.7B-Instruct-v1.0' --gpu-num $NUM_GPUS
diff --git a/python/llm/example/GPU/Pipeline-Parallel-Inference/run_vicuna_arc_2_card.sh b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_vicuna_arc_2_card.sh
new file mode 100644
index 00000000..2f50893e
--- /dev/null
+++ b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_vicuna_arc_2_card.sh
@@ -0,0 +1,41 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+ export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run vicuna-7b-v1.3
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+ generate.py --repo-id-or-model-path 'lmsys/vicuna-7b-v1.3' --gpu-num $NUM_GPUS
+
+# To run vicuna-13b-v1.3
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+# generate.py --repo-id-or-model-path 'lmsys/vicuna-13b-v1.3' --gpu-num $NUM_GPUS
+
+# To run vicuna-33b-v1.3
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+# generate.py --repo-id-or-model-path 'lmsys/vicuna-33b-v1.3' --gpu-num $NUM_GPUS
diff --git a/python/llm/example/GPU/Pipeline-Parallel-Inference/run_yi_arc_2_card.sh b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_yi_arc_2_card.sh
new file mode 100644
index 00000000..ff0a83eb
--- /dev/null
+++ b/python/llm/example/GPU/Pipeline-Parallel-Inference/run_yi_arc_2_card.sh
@@ -0,0 +1,37 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+source /opt/intel/oneapi/setvars.sh
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=9090
+export FI_PROVIDER=tcp
+export USE_XETLA=OFF
+export OMP_NUM_THREADS=6
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+ export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+NUM_GPUS=2 # number of used GPU
+
+# To run Yi-6B-Chat
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+ generate.py --repo-id-or-model-path '01-ai/Yi-6B-Chat' --gpu-num $NUM_GPUS
+
+# To run Yi-34B-Chat
+# CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS \
+# generate.py --repo-id-or-model-path '01-ai/Yi-34B-Chat' --gpu-num $NUM_GPUS
diff --git a/python/llm/src/ipex_llm/transformers/pipeline_parallel.py b/python/llm/src/ipex_llm/transformers/pipeline_parallel.py
index b13e39bc..d346d8aa 100644
--- a/python/llm/src/ipex_llm/transformers/pipeline_parallel.py
+++ b/python/llm/src/ipex_llm/transformers/pipeline_parallel.py
@@ -269,7 +269,8 @@ def pipeline_parallel_generate(self,
"make sure that `pad_token_id` is defined.")
next_ids = next_ids * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)
- if isinstance(outputs.past_key_values, tuple) and local_rank != 0:
+ # Temporarily specify as Baichuan and ChatGLM
+ if self.config.model_type in ["baichuan", "chatglm"] and local_rank != 0:
value_placeholder = torch.empty_like((outputs.past_key_values)[-1][0])
past_key_values_placeholder = tuple(
(value_placeholder, value_placeholder) for _ in range(layer_start)